gitwtfhub

wtf is mkb?

shanghai-academy-of-ai-for-science/mkb — explained in plain English

Analysis updated 2026-05-18

34PythonAudience · researcherComplexity · 5/5LicenseSetup · hard

TL;DR

A single AI foundation model that understands and generates scientific data across proteins, RNA, DNA, molecules, weather, and medical images through one shared backbone.

Mindmap

mindmap
  root((MKB))
    What it does
      Understands multiple sciences
      Generates sequences and forecasts
      Segments medical images
    Tech stack
      Python
      PyTorch
      Qwen3-VL backbone
      CUDA
    Use cases
      Analyze protein or RNA sequences
      Forecast global weather
      Segment medical scan images
    Audience
      Researchers
      AI for science teams
    Requirements
      High end NVIDIA GPU
      Python 3.10
      Downloaded model weights

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

Why would anyone build with this?

REASON 1

Analyze protein, RNA, or DNA sequences with one shared scientific AI model

REASON 2

Design new RNA sequences or generate molecule structures from text descriptions

REASON 3

Forecast global weather ten days ahead using the model's Earth science mode

REASON 4

Segment medical scan images like CT, MRI, or X-ray using text prompts

What's in the stack?

PythonPyTorchQwen3-VLCUDAtransformers

How it stacks up

shanghai-academy-of-ai-for-science/mkbabhisumatk/epstein_files_ragasdfo123/forgewm
Stars343434
LanguagePythonPythonPython
Setup difficultyhardmoderatehard
Complexity5/53/55/5
Audienceresearcherresearcherresearcher

Figures from each repo's GitHub metadata at analysis time.

How do you spin it up?

Difficulty · hard Time to first run · 1day+

Needs an NVIDIA GPU with at least 48GB of memory, a pinned transformers version, and separately downloaded model weights.

Mostly permissive Apache 2.0 terms, but the medical image segmentation part follows Meta's separate SAM license with its own restrictions.

Wtf does this do

MKB, also called Monkey King Bang, is a single AI model built to understand and generate scientific data across many different fields at once, including Earth weather patterns, proteins, RNA, DNA, and small molecules. Instead of needing a separate specialized model for each of these areas, it combines them all into one shared foundation built around an existing large language model called Qwen3-VL, with about 11 billion parameters in total once its extra scientific components are added in. The key design idea is that each type of scientific data keeps its own specialized reader and writer, tuned to that data's particular shape, while all of them share the same central AI backbone. For example, protein sequences use a dedicated protein reader called ESM-2, molecules use a graph based reader, weather data uses a component built for gridded map style data, and medical images use a component based on Meta's SAM image segmentation system. Each of these can also produce output in its native form, whether that is a classification, a designed sequence, a chemical structure written as text, a weather forecast, or a medical image segmentation mask, all accessed through plain natural language instructions rather than separate specialized interfaces for each science. The README backs this up with detailed benchmark comparisons against other scientific AI models, and reports that MKB matches or beats a model roughly 100 times larger on about half of twenty biological understanding tests, while also outperforming or closely matching a leading physics based weather forecasting system used operationally by European forecasters, and matching or beating specialized medical image segmentation tools across most of nine types of medical imaging. Running it yourself is demanding: it needs Python 3.10, an NVIDIA graphics card with at least 48 gigabytes of memory, and a specific CUDA and library setup, with an optional faster attention mode you can compile separately if you want extra speed. Model weights are downloaded separately after installing the required Python packages. It is written in Python and released under a combination of the Apache 2.0 license and Meta's SAM license, reflecting that it reuses parts of Meta's segmentation model.

Yoink these prompts

Prompt 1
Help me set up the MKB conda environment and install the correct PyTorch and CUDA versions.
Prompt 2
Explain how MKB uses separate encoders for protein, RNA, molecule, and weather data within one backbone.
Prompt 3
Show me how to download MKB's model weights and load them for inference.
Prompt 4
Summarize MKB's benchmark results compared to Biology-Instructions and Intern-S1-Pro.

Frequently asked questions

wtf is mkb?

A single AI foundation model that understands and generates scientific data across proteins, RNA, DNA, molecules, weather, and medical images through one shared backbone.

What language is mkb written in?

Mainly Python. The stack also includes Python, PyTorch, Qwen3-VL.

What license does mkb use?

Mostly permissive Apache 2.0 terms, but the medical image segmentation part follows Meta's separate SAM license with its own restrictions.

How hard is mkb to set up?

Setup difficulty is rated hard, with roughly 1day+ to a first successful run.

Who is mkb for?

Mainly researcher.

View the repo → Decode another repo

This repo across BitVibe Labs

Don't trust strangers blindly. Verify against the repo.